A personalized fault diagnosis customization method and system

Abstract

The present invention relates to the field of fault diagnosis and health management, and discloses a personalized fault diagnosis customization method and system. In the initial stage, the method uses a transfer learning strategy to efficiently train the fault diagnosis model of each device with fewer labeled samples. For the subsequent personalized customization process, by considering the factors that may affect the performance of the fault diagnosis model due to the changes in the device's own state and the environment during the vehicle line network information interaction, a personalized customization trigger rule is designed, and a trigger is designed according to these rules. The trigger can quickly respond to the changes in internal and external conditions, trigger the hierarchical fine-tuning of the fault diagnosis model of the device whose performance has changed, so as to cope with the impact of conditions on the diagnosis performance, and help improve the reliability and safety of the device. The present invention effectively improves the accuracy of fault diagnosis of devices with different models and numbers of rail transit vehicles, and has broad application prospects.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis and health management, and particularly to a personalized fault diagnosis customization method and system. Technical Background

[0002] With the rapid development of rail transit, there are many types of rail transit vehicles. The number of vehicles of different types, the number of on-vehicle devices, and the complexity are increasing day by day. Traditional fault diagnosis methods for a single device type without considering the influence of environmental changes on device performance factors are difficult to meet the needs of diverse and complex changes in the vehicle line network conditions. Traditional fault diagnosis usually constructs and generates a fixed-parameter model based on offline data training, lacking the adaptability to changes in the device's own state, such as device aging, component defects, device performance degradation, etc., and changes in the actual working environment of the vehicle line network (rail transit road network, lines, and vehicles actually running on the lines) that cause device performance degradation or faults. Different vehicle types, different devices, running on different lines, or even the same vehicle type and the same model of device running on the same line, different environmental changes (such as sunny, rainy, windy, snowy) may cause different changes in the device's own state, resulting in a significant decline in the diagnostic generalization ability of traditional fault diagnosis methods based on fixed-parameter models. Therefore, there is an urgent need for a personalized fault diagnosis method and system that can customize by considering the factors of changes in the device's own state and environmental changes affecting the performance of the fault diagnosis model in the vehicle line network information interaction. Summary of the Invention

[0003] The purpose of the present invention is to disclose a personalized fault diagnosis customization method and system to achieve personalized customization of the fault diagnosis model.

[0004] To achieve the above object, in a first aspect, the present invention discloses a personalized fault diagnosis customization method, including:

[0005] S1: Collect sensor data of devices with different models and different numbers of rail transit vehicles to obtain a number set and a model set, and construct a mapping relationship from the number set to the model set;

[0006] S2: Based on the mapping relationship, construct a relevant set of the source domain dataset of sensor data, and construct a relevant set of the target source domain dataset of sensor data;

[0007] S3: Initialize the fault diagnosis model for each device, and train the fault diagnosis models of all devices based on the relevant set of the source domain dataset and the relevant set of the target source domain dataset to obtain the parameter-optimal basic fault diagnosis models of all devices;

[0008] S4: Obtain the newly collected sensor data set of the device and its corresponding fault category attribute label set in the vehicle line network information interaction, collect the target semantic information, and construct a personalized customization trigger rule coding matrix based on the target semantic information;

[0009] S5: Construct an optimal decision tree model of the personalized customization trigger rule coding matrix with respect to the set of optimal sub-matrices;

[0010] S6: Construct a hierarchical freezing model based on the parameter-optimal basic fault diagnosis model of all devices;

[0011] S7: Use the personalized customization trigger rule coding matrix, the newly collected sensor data set and its corresponding fault category attribute label set to perform hierarchical fine-tuning on the hierarchical freezing model to construct a personalized customization model.

[0012] To achieve the above object, in a second aspect, the present invention also discloses a personalized fault diagnosis customization method and system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method described in the first aspect above are implemented.

[0013] The present invention has the following beneficial effects

[0014] The personalized fault diagnosis customization method provided by this application uses a transfer learning strategy in the initial stage to efficiently train the fault diagnosis model of each device with fewer labeled samples. For the subsequent personalized customization process, by considering the factors that may affect the performance of the fault diagnosis model due to the change of the device's own state and the environment in the vehicle line network information interaction, personalized customization trigger rules are designed, and triggers are designed according to these rules. This trigger can quickly respond to changes in internal and external conditions, trigger hierarchical fine-tuning of the fault diagnosis model of the device with changed performance, so as to cope with the impact of conditions on the diagnosis performance, can effectively adapt to specific application scenarios and user needs, improve the accuracy and generalization ability of fault diagnosis of devices with different models and numbers of rail transit vehicles, and thus improve the operation reliability and safety of rail transit vehicles. Description of the Drawings

[0015] Figure 1 It is a flowchart of a personalized fault diagnosis customization method according to a preferred embodiment of the present invention;

[0016] Figure 2 It is a visualization result diagram of the source domain and target domain data features extracted by the feature extractor unit when the personalized fault diagnosis customization method according to a preferred embodiment of the present invention uses a transfer learning strategy in the initial stage;

[0017] Figure 3A personalized fault diagnosis customization method according to a preferred embodiment of the present invention during the (k - 1)-th construction of a personalized customization model of the hierarchical freezing model Schematic diagram;

[0018] Figure 4 A personalized fault diagnosis customization method according to a preferred embodiment of the present invention during the (k - 1)-th construction of a personalized customization model Confusion matrix of the model evaluation results before hierarchical fine-tuning;

[0019] Figure 5 A personalized fault diagnosis customization method according to a preferred embodiment of the present invention during the (k - 1)-th construction of a personalized customization model Confusion matrix of the model evaluation results after hierarchical fine-tuning; Detailed implementation manners

[0020] The technical solutions of the present invention will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0021] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, the terms such as "a" or "one" do not indicate a quantity limitation, but mean that there is at least one. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0022] It is worth noting that a personalized fault diagnosis customization method provided in the present application is mainly for the rail transit field and can be applied to rail transit vehicles including high-speed trains / motor trains, regular-speed trains, subway trains, light rail trains, monorail trains, maglev trains, heavy-haul locomotives, etc. Maintenance, inspection and repair stations for these vehicles, such as the inspection and repair stations for high-speed trains / motor trains are train depots and motor train maintenance bases.

[0023] As Figure 1 shown, this embodiment provides a personalized fault diagnosis customization method, including:

[0024] S1: Collect the sensor data of devices with different models and numbers in the rail transit vehicle to obtain a number set and a model set, and construct a mapping relationship from the number set to the model set;

[0025] S2: Based on the mapping relationship, construct a relevant set of the source domain dataset of the sensor data, and construct a relevant set of the target source domain dataset of the sensor data;

[0026] S3: Initialize the fault diagnosis model for each device, and train the fault diagnosis models of all devices based on the relevant set of the source domain dataset and the relevant set of the target source domain dataset to obtain the parameter-optimal basic fault diagnosis models of all devices;

[0027] S4: Obtain the newly collected sensor dataset of the device and its corresponding fault category attribute label set in the vehicle line network information interaction, and collect the target semantic information, and construct a personalized customized trigger rule coding matrix based on the target semantic information;

[0028] S5: Construct an optimal decision tree model of the personalized customized trigger rule coding matrix with respect to the optimal submatrix set;

[0029] S6: Based on the parameter-optimal basic fault diagnosis models of all devices, construct a hierarchical freezing model;

[0030] S7: Use the personalized customized trigger rule coding matrix, the optimal decision tree model, the newly collected sensor dataset and its corresponding fault category attribute label set to perform hierarchical fine-tuning on the hierarchical freezing model to construct a personalized customized model.

[0031] In this embodiment, the device is a device in the rail transit vehicle such as a traction motor. Taking x vehicles under the jurisdiction of a motor car depot as an example, there are y traction motors in each vehicle, then the total number of traction motors is B = xy, that is, there are a total of B numbered devices (traction motors). These B numbered traction motors belong to different models (different model vehicles). For example, devices numbered 1-3 belong to model 1 (vehicle 1), devices numbered 4-6 belong to model 2 (vehicle 2), devices numbered 7-9 belong to model 3 (vehicle 3), devices numbered 10-12 belong to model 3 (vehicle 4), and so on; Thus, the device number b uniquely identifies a certain device of a certain vehicle. For example, b = 10 identifies the first device of vehicle 4, model 3.

[0032] The personalized customization in this application is to customize a fault diagnosis model for each device on each vehicle. The parameter-optimal basic fault diagnosis models for all B devices are constructed based on historical data. For example, if the self-states of devices 1 and 5 change during the k-th vehicle line network information interaction, and the change in the environment affects device 10 to malfunction, then the data collected in the k-th time is used for the k-th personalized customization, that is, only the fault diagnosis model parameters of devices 1 (belonging to model 1), 5 (belonging to model 2), and 10 (belonging to model 3) are finely tuned (each device finely tunes different parameters according to different changes), and the model parameters of other devices 2, 3, 4, 6, 7, 8, 9, 11, and 12 will be frozen and still use the fault diagnosis model customized in the (k - 1)-th time.

[0033] Among them, the sensor data is mainly time series such as the signal of the traction motor stator current sensor. The source domain dataset refers to the labeled dataset, and the target domain dataset refers to the unlabeled dataset. Specifically, both the source domain dataset and the target domain dataset are constructed from the original three-phase current data collected by sensors and are subsets of the original dataset.

[0034] The above-mentioned personalized fault diagnosis customization method uses a transfer learning strategy in the initial stage to efficiently train the fault diagnosis model of each device with fewer labeled samples. For the subsequent personalized customization process, by considering the factors that the change of the device's own state and the environmental change in the vehicle line network information interaction may affect the performance of the fault diagnosis model, a personalized customization trigger rule is designed, and a trigger is designed according to these rules. This trigger can quickly respond to the changes of internal and external conditions, trigger the hierarchical fine-tuning of the fault diagnosis model of the device with performance changes, so as to cope with the impact of conditions on the diagnosis performance, can effectively adapt to specific application scenarios and user needs, improve the accuracy and generalization ability of fault diagnosis for different models and different numbered devices of rail transit vehicles, and further improve the operation reliability and safety of rail transit vehicles.

[0035] It should be noted that the device mentioned in this embodiment is a device in a rail transit vehicle such as a traction motor, and the sensor data collected is the traction motor stator current signal. The main hyperparameters for training the fault diagnosis model in this implementation are shown in Table 1.

[0036] Table 1 Hyperparameters of the basic fault diagnosis model

[0037] Learning rate Number of iterations Batch size Optimizer Learning rate adjustment strategy 0.01 200 64 SGD Cosine annealing

[0038] Furthermore, based on the above parameters, the steps of the personalized fault diagnosis customization method provided in this application are described in detail as follows:

[0039] S11: Collect the sensor data of B numbered devices of A different models of rail transit vehicles to form the original dataset D = {D1,..., Db ,…, D B}, denotes the number of sampling points of the sensor signal, i.e., the total number of elements in the data set D b where b is the device number, b = 1, 2, …, B, and B is the total number of numbers; all device numbers form the number set C B ={1, 2, …, b, … B}; all device models form the model set C A ={1, 2, …, a, … A}, where a is the device model, a = 1, 2, …, A, and A is the total number of models;

[0040] Construct a mapping relationship from the number set to the model set, described as:

[0041] φ: C B → C A ; (1)

[0042] where φ: C B → C A represents the mapping from the number set C B to the model set C A ;

[0043] If the device numbered b belongs to the a-th model, it is denoted as:

[0044] φ(b) = a; a = 1, 2, …, A; (2) [[ID=4l]]

[0045] Construct an index set of all device numbers belonging to the a-th model, denoted as:

[0046] μ a ={b|φ(b) = a, b ∈ {1, 2, …, B}}; (3)

[0047] In the formula, μ a represents the index set of all device numbers belonging to the a-th model;

[0048] Take the first index in μ a to represent the number of the first device belonging to the a-th model device, and min() represents the operation of taking the minimum of all numbers in the parentheses; in the original data set D, take out the corresponding data set a = 1, 2, …, A; construct the source domain data set whose index set is In the formula, denotes taking the first index in μ A ; to represent the number of the first device belonging to the a-th model device; ​

[0049] Label all elements in the source domain dataset with fault category labels, and construct a fault category label set corresponding to the source domain dataset, described as:

[0050]

[0051] In the formula, represents the fault category label set corresponding to the source domain dataset of the a-th model, a = 1, 2,..., A, represents the real number field, represents the fault category label value of the -th element in the source domain dataset of the a-th model, γ z is the z-th label in the fault category attribute label set , Z is the total number of labels, z = ¹, 2,..., Z,

[0052] Construct an index set of all device numbers of the a-th model excluding the first index, denoted as:

[0053]

[0054] In the formula represents the index set of all device numbers of the a-th model excluding the first index arranged in ascending order, x a = 1, 2,..., A a , A a represents the total number of elements in ; Take out the corresponding all indices in x a = 1, 2,..., A a from the original dataset D to construct the target domain dataset:

[0055]

[0056] represents the dataset corresponding to the device numbered ;

[0057] Its index set is:

[0058]

[0059] In the formula, represents the A-th of the A-th model device​​A The number of a device;

[0060] The original data set consists of a source domain data set and a target domain data set Its index set is Optionally, the S3 includes:

[0061] S31: Construct a fault diagnosis model for the device numbered b, described as:

[0062]

[0063] Where represents the fault diagnosis model of the device numbered b, and the symbol represents the function composition operation, respectively represent the feature extractor model and the classifier model of the fault diagnosis model of the device numbered b, and the formulas are respectively:

[0064]

[0065] Where, respectively represent the basic unit of the feature extractor and the basic unit of the classifier of the fault diagnosis model of the device numbered b, θ = 1, 2,..., β, ω = 1, 2,..., α, and β and α respectively represent the number of layers of the feature extractor and the number of layers of the classifier.

[0066] As a preferred implementation manner of this embodiment, the feature extractor model of the fault diagnosis model is an 18-layer one-dimensional residual neural network, which includes a total of 5 basic units, β = 5, and the basic unit of the feature extractor includes 1 one-dimensional convolution, 1 batch normalization layer and 1 max pooling layer. The basic unit of the feature extractor ~basic unit Each basic unit only includes 1 residual block. The basic unit of the feature extractor includes 1 residual block, 1 average pooling layer and 1 flattening layer; the classifier model of the fault diagnosis model is a multi-layer perceptron network, including 1 basic unit of the classifier α = 1 This basic unit of the classifier includes 1 linear layer.

[0067] S32: Use the source domain and target domain data sets to train the fault diagnosis models of all devices to obtain the parameter-optimal basic fault diagnosis models of all devices.

[0068] For the devices indicated by the indexes in the index set belonging to the a-th model, b = 1, 2,..., B, a = 1, 2,..., A, use the source domain data set index set All the indexes in the dataset referred to Train a fault diagnosis model Optimize and update it using the gradient descent method to obtain the parameter-optimal basic fault diagnosis model of the device referred to by the index in the index set belonging to the a-th model in the index set Its model parameter optimization function is expressed as:

[0069]

[0070] Among them, represents the optimal model parameters of the parameter-optimal basic fault diagnosis model of the device belonging to the a-th model and numbered the parameter-optimal basic fault diagnosis model of the device of; represents the model parameters of the fault diagnosis model of the device belonging to the a-th model and numbered the device of; represents the parameters when taking the minimum condition of the operation result in the parentheses among all parameters L L ce (·) represents the cross-entropy loss function; represents that the model parameters are when the input is the source domain dataset the output of the fault diagnosis model γ a (·) represents the label mapping function of the a-type device.

[0071] Construct the set of all parameter-optimal basic fault diagnosis models of the devices referred to by the indexes in the index set belonging to the a-th model in the index set Among them, the parameter-optimal basic fault diagnosis model of the device numbered is expressed as: is expressed as:

[0072] (10)

[0073] In the formula respectively represent the optimal feature extractor model and classifier model of the parameter-optimal basic fault diagnosis model The formulas are respectively:

[0074]

[0075] Among them, respectively represent the basic unit of the optimal feature extractor and the basic unit of the classifier.

[0076] For the devices belonging to the a-th model and numbered in the index set The device pointed to by the index in is b=1,2,…,B, a=1,2,…,A, using the target domain dataset Index Collection All indexes in x a =1,2,…,A a , refers to the target domain dataset Training a fault diagnosis model Get the index set belonging to the ath model and numbered Optimal basic fault diagnosis model for the device pointed to by the index in The model parameter optimization function is expressed as:

[0077]

[0078] in, Indicates that it belongs to the ath model and is numbered as an index set Middle Index Parameter-optimal basic fault diagnosis model for the indicated equipment The optimal model parameters are Indicates that it belongs to the ath model and is numbered as an index set Middle Index Fault diagnosis model of the device in question Model parameters of Indicates that all parameters Take the parameter that satisfies the minimum condition of the operation result in the brackets L adapt (·) represents the domain adaptation loss function, η∈[0,1] represents the weight parameter of the domain adaptation loss; The model parameters are Input is the source domain dataset Real-time fault diagnosis model The output, Respectively represent the input as the target domain dataset Source domain dataset Real-time fault diagnosis model Feature extractor model Output; In this embodiment, two groups of asynchronous AC motors with the same model and different working conditions are selected for transfer learning, where the domain adaptation loss function selects the maximum mean difference loss, η is 0.5, and the visualization results of the source domain and target domain data features extracted by the feature extractor unit are shown in the figure Figure 2 As shown;

[0079] Construct an index set belonging to the ath model and numbered Optimal basic fault diagnosis model for all parameters of the device pointed to by the middle index be expressed as: be expressed as:

[0080]

[0081] In the formula respectively represent the optimal feature extractor model and classifier model of the optimal basic fault diagnosis model of the parameters, and the formulas are respectively: be expressed as:

[0082]

[0083] Among them, respectively represent the optimal feature extractor basic unit and classifier basic unit of the optimal basic fault diagnosis model of the parameters, x a = 1, 2, …, A a , a = 1, 2, …, A.

[0084] From formulas (9) to (16), the set of optimal basic fault diagnosis models of all devices' parameters is expressed as: is expressed as:

[0085]

[0086] The said S4 includes:

[0087] S41: Collect the newly acquired sensor data set of the device and its corresponding label set; collect the semantic information required to construct the personalized customized trigger rule coding matrix;

[0088] Collect the sensor data set in the vehicle line network information interaction collected by the device numbered b (b = 1, 2, 3, …, B) for the kth time and its corresponding fault category attribute label set γ z is the zth label in the fault category attribute label set , z = 1, 2, …, Z, Z is the total number of labels, k is the number of personalized customizations, k = 1, 2, 3, …, K, K is the total number of personalized customizations; collect the semantic information in the vehicle line network information interaction required to construct the personalized customized trigger rule coding matrix for the kth time;

[0089] S42: Construct the personalized customized trigger rule coding matrix.

[0090] Encode the semantic information affecting the trigger result in the collected semantic information for constructing the personalized customized trigger rule coding matrix for the kth time, and construct the kth personalized customized trigger rule coding matrix, the formula is:

[0091]

[0091]

[0092] in, The personalized trigger rule encoding matrix constructed for the kth time, Represents the personalized customization trigger rule encoding matrix P k The element at row i and column j, i = 1, 2, 3, ..., n, j = 1, 2, 3, ..., m, n is the personalized customized trigger rule encoding matrix P constructed for the kth time k The total number of rules (total number of rows), m is the total number of semantic information encodings (total number of columns) of the personalized customization trigger rule encoding matrix constructed for the kth time.

[0093] The S5 includes:

[0094] Construct an optimal decision tree model of the optimal sub-matrix set of the personalized customized trigger rule encoding matrix.

[0095] S51: Calculate the entropy of the k-th construction of the personalized customization trigger rule encoding matrix.

[0096] Construct the row rule trigger result decision column vector corresponding to the k-th personalized customization trigger rule encoding matrix. The formula is:

[0097]

[0098] In the formula Represents the kth personalized customization trigger rule coding matrix The corresponding row rule trigger result decision vector; o k；i P k The i-th row rule The corresponding trigger result, o k；i ∈{0,1}, i=1,2,3,…,n;

[0099] Calculate the entropy of the encoding matrix of the k-th personalized customization trigger rule. The formula is:

[0100]

[0101] Among them, H(P k ) is the kth personalized customization trigger rule encoding matrix The entropy of Respectively represent P k The rule in row i Corresponding trigger result o k；i (i=1,2,3,…,n) The number of rows with 0 and 1 accounts for P k The ratio of the total number of rows is:

[0102]

[0103] where 1(·) is an indicator function that takes the value 1 when the condition in the parentheses is true and 0 otherwise;

[0104] S52: Calculate the entropy of the sub-matrix of the k-th personalized customization trigger rule coding matrix.

[0105] Construct the sub-matrix of the k-th personalized customization trigger rule coding matrix, with the formula:

[0106]

[0107] where is the j-th column element of the k-th personalized customization trigger rule coding matrix and the sub-matrix with the value v, where v ∈ {0, 1} represents the personalized customization trigger rule coding matrix P at the i-th (i = 1, 2, 3,..., n) row and j-th (j = 1, 2, 3,..., m) column element k and the value of is and P k is the j-th column element of and the number of rows of the sub-matrix with the value v is where i ∈ {1, 2,..., n}} is the row set of the sub-matrix, and size(·) represents the number of elements in the set, that is, the number of rows of the sub-matrix. Specifically, when the column elements take the values

[0108] respectively, construct the sub-matrices of the k-th personalized customization trigger rule coding matrix as and and are respectively the number of rows of the sub-matrix when the column elements take the values and is

[0109] Construct the j-th (j = 1, 2, 3,..., m) column element k of the k-th personalized customization trigger rule coding matrix P with all sub-matrices where the values are v ∈ {0, 1} respectively as: The number of rows corresponding to all sub-matrices is:

[0110] Calculate the entropy of the sub-matrix of the k-th personalized customization trigger rule coding matrix, with the formula:

[0111]

[0112] where, is the k-th personalized customization trigger rule coding matrix The j-th column element of The sub-matrix with values v ∈ {0, 1} The entropy of Respectively represent the sub-matrices The row rule trigger results in The ratio of the number of rows with values 0 and 1 to the total number of rows in the sub-matrix The formula is:

[0113]

[0114] The k-th personalized customization trigger rule coding matrix P k The j (j = 1, 2, 3, …, m)-th column element of The entropies of all sub-matrices with values v ∈ {0, 1} respectively are:

[0115] S53: Calculate the column where the column element with the largest information gain in the personalized customization trigger rule coding matrix is located;

[0116] Calculate the k-th personalized customization trigger rule coding matrix P k The j-th column element of The information gain of, the formula is:

[0117]

[0118] In the formula Is the information gain of the j-th column element of the personalized customization trigger rule coding matrix P k The j-th column element of The information gain;

[0119] The k-th personalized customization trigger rule coding matrix P k The j (j = 1, 2, 3, …, m)-th column element of All information gains are

[0120] Select the column element with the largest information gain Obtain the column where it is located, the formula is:

[0121]

[0122] In the formula, j* is the column element with the largest information gain in the personalized customization trigger rule coding matrix P k The column where it is located, j ∈ {1, 2, 3, …, m}, * Indicates taking the column j (j = 1, 2, 3, …, m) when the result of the operation in the parentheses is the largest;

[0123] ​S54: Construct the k-th personalized customization trigger rule coding matrix P according to the column where the element with the maximum information gain is located k of the L-th set of optimal submatrix sets;

[0124] According to the personalized customization trigger rule coding matrix P k where the maximum information gain is located in the j-th * column element take values v ∈ {0, 1} respectively to construct the first set of optimal submatrices That is and are the row numbers of the optimal submatrices when the column elements take values respectively, The number of the first set of optimal submatrices is N1 = 2 1 = 2;

[0125] For the first set of optimal submatrices k in the personalized customization trigger rule coding matrix P according to the column where the maximum information gain is located column element take values v ∈ {0, 1} respectively to construct the second set of optimal submatrices That is and are the second set of optimal submatrices, are the row numbers of the optimal submatrices respectively, and The number of the second set of optimal submatrices is N2 = 2 2 = 4; (The submatrix can be an empty set).

[0126] For the (l - 1)-th set of optimal submatrices k in the personalized customization trigger rule coding matrix P according to the column where the maximum information gain is located column element take values v ∈ {0, 1} respectively to construct the l-th set of optimal submatrices The number of the l-th set of optimal submatrices is N l = 2 l ;

[0127] For the (L - 1)-th set of optimal submatrices k in the personalized customization trigger rule coding matrix P according to the column where the maximum information gain is located column element take values v ∈ {0, 1} respectively to construct the L-th set of optimal submatrix sets The number of the L-th set of optimal submatrices is N L = 2 L; At this time, the personalized customization trigger rule coding matrix P k in the entropy of each optimal sub-matrix in the L-th group of optimal sub-matrix sets is equal to 0; the personalized customization trigger rule coding matrix P k in the L-th group of optimal sub-matrix sets is expressed as:

[0128]

[0129] wherein, is the t-th optimal sub-matrix in the L-th group of optimal sub-matrix sets of the personalized customization trigger rule coding matrix P k in the L-th group of optimal sub-matrix sets where t = 1, 2, …, T, and T = N L = 2 L ; wherein, The brackets of represent another concise expression of the symbol before the brackets, with the same meaning; see the left and right sides of the first equal sign in Formula 27. The personalized customization trigger rule coding matrix P k the L-th group of optimal sub-matrix sets the t-th sub-matrix in the rule of the i-th row the trigger result of is equal to c t ∈ {0, 1}, i ∈ {1, 2, …, n}} is the row set of the t-th optimal sub-matrix , is the number of rows of the t-th optimal sub-matrix , the personalized customization trigger rule coding matrix P k in the L-th group of optimal sub-matrix sets the trigger result set is C = {c1, c2, …, c t , …, c T}.

[0130] S55: Construct the optimal decision tree model of the personalized customization trigger rule coding matrix with respect to the L-th group of optimal sub-matrix sets. The formula is:

[0131]

[0132] wherein, is the optimal decision tree model of the personalized customization trigger rule coding matrix P k with respect to the L-th group of optimal sub-matrix sets ; means taking any i-th row from P k as the input of the decision tree model, i = 1, 2, 3, …, n; c t is the trigger result of the t-th optimal sub-matrix, Indicates that the indicator function takes the value of 1 when the condition within the parentheses holds.

[0133] S6 includes:

[0134] Based on the set M of optimal basic fault diagnosis models for all device parameters b , construct a hierarchical freezing model, and the formula is:

[0135]

[0136] where represents the hierarchical freezing model of the device numbered b, represents the model M of the device numbered b b 's freezing operation function, are respectively the set of indicator functions for whether each layer in

[0137] are respectively the indicator functions for whether the layers in are frozen. The value 0 indicates freezing (in the frozen state, the input of this layer is directly passed to the output without updating the parameters of this layer), and the value 1 indicates activation (in the active state, the parameters of this layer are updated); respectively represent the feature extractor model and the classifier model of the hierarchical freezing model of the device numbered b, respectively represent the basic unit of the feature extractor and the basic unit of the classifier of the hierarchical freezing model of the device numbered b, θ = 1, 2,..., β, ω = 1, 2,..., α, and β and α respectively represent the number of layers of the feature extractor and the number of layers of the classifier,

[0138]

[0139] where id(·) is the identity function for the basic unit within the parentheses, used to activate the update of the parameters of the corresponding basic unit; respectively represent the basic unit of the feature extractor and the basic unit of the classifier of the optimal basic fault diagnosis model of the device numbered b; b = 1, 2, 3,..., B.

[0140] S7 includes:

[0141] S71: Utilize the constructed k-th personalized customization trigger rule coding matrix P k , and according to formula (28), calculate the optimal decision result k of the k-th personalized customization trigger rule coding matrix P with respect to the set of optimal sub-matrices of the L-th group The row trigger rule determined when constructing the k-th personalized customization model i = 1, 2, 3, …, n;

[0142] When it is the case, trigger the (k - 1)-th personalized customization model to perform the k-th hierarchical fine-tuning. Utilize the sensor data set newly collected for the k-th time by devices numbered b (b = 1, 2, 3, …, B) and its corresponding set of fault category attribute labels to construct the k-th personalized customization model. The formula is:

[0143]

[0144] where respectively represent the personalized customization models constructed for the k-th and (k - 1)-th times by the device numbered b represents the fine-tuning operation function of the personalized customization model constructed for the k-th time by the device numbered b represents the fine-tuning learning rate of the personalized customization model constructed for the k-th time by the device numbered b; respectively represent the personalized customization model constructed for the k-th time by the device numbered b 's feature extractor model and classifier model; respectively represent the personalized customization model constructed for the k-th time by the device numbered b 's basic unit of the feature extractor and basic unit of the classifier, θ = 1, 2, …, β, ω = 1, 2, …, α, and β and α respectively represent the number of layers of the feature extractor and the number of layers of the classifier; represents the personalized customization model constructed for the (k - 1)-th time 's freezing operation function, satisfying the following relational expression:

[0145]

[0146] In the formula, represents the hierarchical freezing model of the personalized customization model constructed for the (k - 1)-th time by the device numbered b respectively represent the hierarchical freezing model of the personalized customization model constructed for the (k - 1)-th time by the device numbered b 's feature extractor model and classifier model of the hierarchical freezing model respectively satisfy the following expressions: respectively represent the personalized customization model constructed for the (k - 1)-th time by the device numbered b of the hierarchical freezing model 's basic unit of the feature extractor and basic unit of the classifier, respectively satisfy the following expressions:

[0147]

[0148] where id(·) is the identity function for the basic unit within the parentheses, used to activate the update of the parameters of the corresponding basic unit; respectively represent the personalized customization model constructed for the (k - 1)-th time by device numbered b feature extractor model classifier model indicator function indicating whether the layers in are frozen, with a value of 0 indicating frozen and a value of 1 indicating activated; respectively represent the personalized customization model constructed for the (k - 1)-th time by device numbered b feature extractor basic unit, classifier basic unit of;

[0149] Specifically, when k = 1, it represents the personalized customization model U constructed for the 0-th time by device numbered b b 0 hierarchical freezing model of is its freezing operation function is where that is from the hierarchical freezing model of the personalized customization model constructed for the 0-th time hierarchical freezing model of perform the first fine-tuning on to obtain the first personalized customization model constructed as

[0150] When , freeze (do not trigger) the (k - 1)-th personalized customization model perform the k-th hierarchical fine-tuning, and let the k-th constructed personalized customization model be

[0151] Let k = k + 1, k = 1, 2, 3,..., K, go to execute S21, and stop the operation until k > K.

[0152] Specifically in this embodiment, the hierarchical freezing model of the personalized customization model constructed for the (k - 1)-th time by device numbered b is schematic diagram as Figure 3 shown, the feature extractor basic units 1, 2, 3 are frozen, the feature extractor basic units 4, 5 and the classifier basic unit are activated, receive the k-th vehicle lane network interaction information, and according to the decision results of the decision tree model of the personalized customization trigger rule, the confusion matrices of the model evaluation results before and after the k-th hierarchical fine-tuning are respectively as Figure 4 、 5 shown.

[0153] In summary, the personalized fault diagnosis customization method provided in this application uses a transfer learning strategy in the initial stage to efficiently train the fault diagnosis model of each device with fewer labeled samples. For the subsequent personalized customization process, by considering the factors that may affect the performance of the fault diagnosis model due to the changes in the device's own state and the environment during the vehicle line network information interaction, a personalized customization trigger rule can be designed, and a trigger can be designed according to these rules. Through the personalized rule trigger to interact with the outside world, the adjustment of the fault diagnosis model of the device with changed performance is triggered, so as to cope with the impact of conditions on the diagnosis performance. The implementation of this method can be applied to various existing devices without changing the existing hardware devices, which is of great significance for improving the reliability of device operation and maintaining system security, etc.

[0154] This application also provides a personalized fault diagnosis customization system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented. This personalized fault diagnosis customization system can implement each embodiment of the above personalized fault diagnosis customization method and achieve the same effect, which will not be elaborated here.

[0155] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should be within the protection scope determined by the claims.

Claims

1. A personalized fault diagnosis customization method, characterized in that Including: S1: Collect sensor data of devices with different models and numbers in rail transit vehicles to obtain a number set and a model set, and construct a mapping relationship from the number set to the model set; S2: Based on the mapping relationship, construct a relevant set of the source domain dataset of sensor data, and construct a relevant set of the target source domain dataset of sensor data; S3: Initialize the fault diagnosis model for each device, and train the fault diagnosis models of all devices based on the relevant set of the source domain dataset and the relevant set of the target source domain dataset to obtain the parameter-optimal basic fault diagnosis models of all devices; S4: Obtain the newly collected sensor dataset of the device and its corresponding fault category attribute label set in the vehicle line network information interaction, and collect target semantic information, and construct a personalized customized trigger rule coding matrix based on the target semantic information; S5: Construct an optimal decision tree model of the personalized customized trigger rule coding matrix with respect to the set of optimal submatrices; S6: Based on the parameter-optimal basic fault diagnosis models of all devices, construct a hierarchical freezing model; S7: Use the personalized customized trigger rule coding matrix, the optimal decision tree model, the newly collected sensor dataset and its corresponding fault category attribute label set to perform hierarchical fine-tuning on the hierarchical freezing model to construct a personalized customized model.

2. The personalized fault diagnosis customization method according to claim 1, characterized in that The S1 includes: S11: Collect the sensor data of B numbered devices of A different models of rail transit vehicles to form an original data set D = {D1, …, D b , …, D B}, represents the number of sampling points of the sensor signal, b is the device number, b = 1, 2, …, B, and B is the total number of numbers; all device numbers form a number set C B = {1, 2, …, b, … B}; all device models form a model set C A = {1, 2, …, a, … A}, a is the device model, a = 1, 2, …, A, and A is the total number of models; Construct a mapping relationship from the number set to the model set, described as follows: φ: C B → C A ; (1) Among them, φ: C B → C A represents the mapping from the number set C B to the model set C A .

3. The personalized fault diagnosis customization method according to claim 1, wherein The S2 includes: If the device with number b belongs to the a-th model, denoted as: φ(b) = a; a = 1, 2, …, A; (2) In the formula, A is the total number of models; Construct an index set of all device numbers belonging to the a-th model, denoted as: μ a = {b | φ(b) = a, b ∈ {1, 2, …, B}}; (3) where μ a represents the index set of all device numbers belonging to the a-th model, and B is the total number of numbers; Take μ a The first index in represents the number of the first device belonging to the a-th type of device, and min() represents the operation of taking the minimum of all numbers within the parentheses; in the original dataset D, the corresponding dataset is taken out a = 1, 2, …, A; construct the source domain dataset Its index set is as follows: wherein, represents taking the first index in μ A , and represents the number of the first device belonging to the first type of device A; Label all elements in the source domain dataset with fault category labels to construct a fault category label set corresponding to the source domain dataset, described as follows: In the formula, represents the set of fault category labels corresponding to the source domain data set of the a-th model, where a = 1, 2, …, A, represents the real number field, represents the fault category label value of the n-th element in the source domain data set of the a-th model, γ z is the z-th label in the set of fault category attribute labels where Z is the total number of labels, and z = 1, 2, …, Z, Construct an index set of other indexes except the first index in the index set of all device numbers belonging to the a-th model, denoted as: In the formula, represents the set of other indices, arranged in ascending order, obtained by removing the first index from the set of all device number indices belonging to the a-th model , x a = 1, 2, …, A a , where A a represents the total number of elements in Retrieve the corresponding ones in the original dataset D All indexes in x a = 1, 2, …, A a of the dataset, where a = 1, 2, …, A, and construct the target domain dataset as follows: In the formula, represents the data set corresponding to the device numbered ; The index set of the target domain dataset is as follows: In the formula, represents the number of the A A-th device belonging to the A-type device; The original dataset consists of a source domain dataset and a target domain dataset Its index set is 4. The personalized fault diagnosis customization method according to claim 1, wherein The S3 includes: S31: Construct a fault diagnosis model with device number b, described as: Among them represents the fault diagnosis model of the device numbered b, and the symbol ο represents the function composition operation respectively represent the feature extractor model and the classifier model of the fault diagnosis model of the device numbered b, and the formulas are respectively Among them, respectively represent the basic unit of the feature extractor and the basic unit of the classifier of the device fault diagnosis model numbered b, where θ = 1, 2, …, β, ω = 1, 2, …, α, and β and α respectively represent the number of layers of the feature extractor and the number of layers of the classifier; S32: Use the relevant set of the source domain dataset and the relevant set of the target source domain dataset to train the fault diagnosis models of all devices to obtain the parameter-optimal basic fault diagnosis models of all devices: For the devices indicated by the indices in the index set belonging to the a-th model type, with b = 1, 2, …, B and a = 1, 2, …, A, using the source domain dataset in the index set index set all indices in the indicated dataset to train a fault diagnosis model Optimize and update using the gradient descent method to obtain the parameter-optimal basic fault diagnosis model for the devices indicated by the indices in the index set belonging to the a-th model type The model parameter optimization function satisfies the following relationship: ​ Among them, represents the optimal basic fault diagnosis model of the device with the a-th model type and the number , and the optimal model parameters ; represents the model parameters of the fault diagnosis model of the device with the a-th model type and the number ; represents the parameter that satisfies the condition of the smallest operation result in the parentheses when taking all parameters ; L ce (·) represents the cross-entropy loss function; represents that the model parameters are , and when the input is the source domain data set , the output of the fault diagnosis model is γ a (·) represents the label mapping function of the device of model a; Construct all-parameter optimal basic fault diagnosis model sets for devices pointed to by indexes in the index set with the a-th model number Among them, the parameter-optimal basic fault diagnosis model of the device numbered is expressed as: is expressed as:​ In the formula, respectively represent the optimal feature extraction model and the optimal classifier model of the parameter-optimal basic fault diagnosis model The formulas are as follows: Among them, respectively represent the basic unit of the optimal feature extractor and the basic unit of the classifier; For the devices indicated by the indices in the index set belonging to the a-th model type, with b = 1, 2, …, B and a = 1, 2, …, A, using the target domain dataset in the index set index set all indices in x a = 1, 2, …, A a , the indicated target domain dataset to train a fault diagnosis model to obtain a parameter-optimal basic fault diagnosis model for the devices indicated by the indices in the index set belonging to the a-th model type where The model parameter optimization function is expressed as follows: ​ Among them, represents the optimal basic fault diagnosis model of the device pointed to by the index in the index set numbered belonging to the a-th model, and the optimal model parameters of the optimal model; represents the model parameters of the fault diagnosis model of the device pointed to by the index in the index set numbered belonging to the a-th model; represents the parameter that satisfies the condition of the smallest operation result in the parentheses among all parameters L adapt (·) represents the domain adaptation loss function, and η ∈ [0, 1] represents the weight parameter of the domain adaptation loss; represents that the model parameter is when the input is the source domain dataset the output of the fault diagnosis model respectively represent the output of the feature extractor model of the fault diagnosis model when the input is the target domain dataset the source domain dataset Construct all-parameter optimal basic fault diagnosis models for devices indicated by indices in the index set of the a-th model type and numbered as follows: Let be expressed as: In the formula, respectively represent the optimal feature extraction model and classifier model of the parameter-optimal basic fault diagnosis model The formulas are respectively as follows: Among them, respectively represent the basic units of the optimal feature extractor and the basic units of the classifier of the optimal basic fault diagnosis model ; x a = 1, 2, …, A a , a = 1, 2, …, A; From formulas (9) to (16), obtain the set of parameter-optimal basic fault diagnosis models for all devices Expressed as:

5. A personalized fault diagnosis customization method according to claim 1, characterized in that The S4 includes: S41: Collect the sensor data set of the k-th collection of vehicle line network information interaction of devices numbered b, where b = 1, 2, 3,..., B and its corresponding set of fault category attribute labels γ z is the z-th label in the set of fault category attribute labels, where z = 1, 2,..., Z, Z is the total number of labels, k is the number of personalized customization times, k = 1, 2, 3,..., K, and K is the total number of personalized customization times; collect the semantic information in the vehicle line network information interaction required for constructing the personalized customization trigger rule coding matrix in the k-th time ​ S42: Encode the semantic information affecting the trigger result in the k-th construction of the personalized customized trigger rule coding matrix, and construct the k-th personalized customized trigger rule coding matrix, the formula is as follows: Among them, is the personalized customization trigger rule coding matrix constructed for the k-th time, indicating the personalized customization trigger rule coding matrix P k The element at the i-th row and j-th column, where i = 1, 2, 3, …, n, j = 1, 2, 3, …, m, and n is the total number of rows of the personalized customization trigger rule coding matrix P k constructed for the k-th time, and m is the total number of columns of the semantic information coding of the personalized customization trigger rule coding matrix constructed for the k-th time.

6. The personalized fault diagnosis customization method according to claim 1, wherein The S5 includes: S51: Construct a row rule trigger result decision column vector corresponding to the k-th personalized customized trigger rule coding matrix, the formula is as follows: O k = [o k；1 o k；2 … o k；i … o k；n T ; (19)​ Wherein, represents the k-th personalized customization trigger rule coding matrix corresponding row rule trigger result decision vector; o k；i is the i-th row rule of P k and o is the corresponding trigger result, o k；i ∈{0,1}, i = 1, 2, 3, …, n; Calculate the entropy of the k-th personalized customized trigger rule coding matrix, the formula is as follows: Among them, H(P k ) is the entropy of the k-th personalized customization trigger rule coding matrix . respectively represent the trigger result o k corresponding to the i-th row rule in P k；i where the number of rows with the value of 0 or 1 accounts for the ratio of the total number of rows of P k . The total number of rows is i = 1, 2, 3,..., n. The formula is as follows: In the formula, 1(·) is an indicator function, which takes the value 1 when the condition in the parentheses is true, otherwise it takes the value 0; S52: Construct a submatrix of the k-th personalized customized trigger rule coding matrix, the formula is as follows: Among them, is the k-th personalized customization trigger rule coding matrix of the j-th column element The sub-matrix with the value of v, where v ∈ {0, 1} represents the personalized customization trigger rule coding matrix P k The element at the i-th row and j-th column value, where i = 1, 2, 3, …, n and j = 1, 2, 3, …, m, is P k The j-th column element in The sub-matrix with the value of v number of rows, i ∈ {1, 2, …, n} is the row set of the sub-matrix, and size(·) represents the number of elements in the set; Construct the k-th personalized customization trigger rule encoding matrix P k The j-th column element of All sub-matrices with elements v ∈ {0, 1} respectively are as follows: The number of rows corresponding to all sub-matrices is: Calculate the entropy of the submatrix of the k-th personalized customized trigger rule coding matrix, the formula is: wherein, is the encoding matrix of the k-th personalized customization trigger rule the j-th column element of the sub-matrix with values v ∈ {0, 1} the entropy of respectively represent the sub-matrix the row rule trigger result o k；i is the ratio of the number of rows with values 0 and 1 in to the total number of rows of the sub-matrix The formula is as follows: The encoding matrix \(P\) of the \(k\)-th personalized customization trigger rule k The \(j\)-th column element The entropies of all submatrices with values \(v\in\{0, 1\}\) are as follows, where \(j = 1, 2, 3,\cdots, m\): S53: Calculate the information gain of the j-th column element of the k-th personalized customization trigger rule encoding matrix P k as follows: The formula is as follows: In the formula, is the personalized customization trigger rule coding matrix P k is the j-th column element of and is the information gain of The encoding matrix P of the k-th personalized customization trigger rule k The j-th column element of All information gains are Select the column element with the largest information gain Obtain the column where it is located. The formula is as follows: where, j* is the column element with the largest information gain in the personalized customization trigger rule coding matrix P k in the column where it is located, j * ∈ {1, 2, 3, …, m}, represents the column j when taking the condition where the operation result in the parentheses is the largest; S234: Construct the k-th personalized customization trigger rule coding matrix P according to the column where the element with the largest information gain is located k of the L-th set of optimal submatrix sets; Encode the personalized customization trigger rule encoding matrix P k The optimal submatrix of the (l - 1)-th group in According to the elements in the column where the information gain is the largest respectively Construct the optimal submatrix of the l-th group by taking values v ∈ {0, 1} for the column elements The number of optimal submatrices of the l-th group is N l l = 2 l ; Encode the personalized customization trigger rule encoding matrix P k The (L - 1)-th group of optimal sub-matrices in According to the columns where their information gains are the largest respectively Elements Take values of v ∈ {0, 1} to construct the set of the L-th group of optimal sub-matrices The number of the L-th group of optimal sub-matrices is N L = 2 L ; At this time, the entropy of each optimal sub-matrix in the set of the L-th group of optimal sub-matrices in the personalized customization trigger rule encoding matrix P k is equal to 0; The set of the L-th group of optimal sub-matrices in the personalized customization trigger rule encoding matrix P k is represented as follows: is shown as follows: Among them, is the personalized customization trigger rule coding matrix P k the t-th optimal submatrix in the L-th set of optimal submatrices where t = 1, 2, …, T, and T = N L = 2 L ; the personalized customization trigger rule coding matrix P k the L-th set of optimal submatrices the t-th submatrix the rule of the i-th row the trigger result equals c t ∈ {0, 1}, i ∈ {1, 2, …, n}} is the row set of the t-th optimal submatrix and is the number of rows of the t-th optimal submatrix ; the personalized customization trigger rule coding matrix P k the L-th set of optimal submatrices the trigger result set is C = {c1, c2, …, c t , …, c T}; S55: Construct an optimal decision tree model of the personalized customized trigger rule coding matrix with respect to the L-th group of optimal submatrix sets, the formula is as follows: Among them, is the encoding matrix P for the personalized customization trigger rule k Regarding the set of optimal sub-matrices of the L-th group of the optimal decision tree model; denotes taking any i-th row from P k as the input of the decision tree model, where i = 1, 2, 3, …, n; c t is the trigger result of the t-th optimal sub-matrix, denotes that the indicator function takes the value of 1 when the condition inside the parentheses holds.

7. The personalized fault diagnosis customization method according to claim 1, wherein The S6 includes: Set M of optimal basic fault diagnosis models based on the parameters of all devices b , construct a hierarchical freezing model, and the formula is as follows: Among them, represents the hierarchical freezing model of the device numbered b, represents the model M of the device numbered b b freezing operation function of are respectively the set of indication functions for whether each layer in are respectively the indication functions for whether the layers in are frozen, where the value 0 indicates freezing and the value 1 indicates activation; respectively represent the feature extractor model and the classifier model of the hierarchical freezing model of the device numbered b, respectively represent the basic unit of the feature extractor and the basic unit of the classifier of the hierarchical freezing model of the device numbered b, θ = 1, 2, …, β, ω = 1, 2, …, α, and β and α respectively represent the number of layers of the feature extractor and the number of layers of the classifier, respectively satisfy the following expressions: Among them, id(·) is the identity function for the basic unit within the parentheses, which is used to activate the update of the parameters of the corresponding basic unit; respectively represent the feature extractor basic unit and the classifier basic unit of the optimal basic fault diagnosis model of the device parameters numbered b; b = 1, 2, 3, …, B.

8. The personalized fault diagnosis customization method according to claim 1, characterized in that The S7 includes: S71: Utilize the personalized customization trigger rule encoding matrix P constructed at the k-th time k , and calculate the personalized customization trigger rule encoding matrix P according to the optimal decision tree model of the personalized customization trigger rule encoding matrix with respect to the optimal sub-matrix set of the L-th group k with respect to the optimal sub-matrix set of the L-th group the optimal decision result is the trigger rule of the row determined when constructing the k-th personalized customization model When it triggers the (k - 1)-th personalized customization model to perform the k-th hierarchical fine-tuning. Using the newly collected sensor data sets of the device numbered b, b = 1, 2, 3, …, B, at the k-th time and their corresponding fault category attribute label sets to construct the k-th personalized customization model, the formula is as follows: Among them, respectively represent the personalized customization models constructed for the b-numbered device at the k-th and (k - 1)-th times, represents the fine-tuning operation function of the personalized customization model constructed for the b-numbered device at the k-th time, represents the fine-tuning learning rate of the personalized customization model constructed for the b-numbered device at the k-th time; respectively represent the personalized customization model constructed for the b-numbered device at the k-th time 's feature extractor model and classifier model; respectively represent the basic units of the feature extractor and the basic units of the classifier of the personalized customization model constructed for the b-numbered device at the k-th time , where θ = 1, 2, …, β, ω = 1, 2, …, α, and β and α respectively represent the number of layers of the feature extractor and the number of layers of the classifier; represents the freezing operation function of the personalized customization model constructed at the (k - 1)-th time satisfies the following relational expression: In the formula, represents the personalized customization model constructed for the (k - 1)-th time by the device numbered b of the hierarchical freezing model, respectively represent the personalized customization model constructed for the (k - 1)-th time by the device numbered b of the hierarchical freezing model of the feature extractor model and the classifier model, respectively represent the personalized customization model constructed for the (k - 1)-th time by the device numbered b of the hierarchical freezing model of the feature extractor basic unit and the classifier basic unit, respectively satisfy the following expressions: where, id(·) is the identity function for the basic unit within the parentheses, which is used to activate the update of the parameters of the corresponding basic unit; respectively represent the personalized customization model constructed for the (k-1)-th time by the device numbered b feature extraction model classifier model is an indicator function for whether the layer in is frozen. A value of 0 indicates freezing, and a value of 1 indicates activation; respectively represent the personalized customization model constructed for the (k-1)-th time by the device numbered b feature extraction basic unit and classifier basic unit of; When freeze the (k - 1)-th personalized customization model perform the k-th hierarchical fine-tuning, and let the personalized customization model constructed in the k-th time be Let \(k = k + 1\), where \(k=1,2,3,\cdots,K\), then go to step S21 for execution, and stop the operation until \(k>K\).

9. A personalized fault diagnosis customization system, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1-8 above.

Images